Computer Vision-Based Home Accident Detection and Response System for Elderly Care

This paper proposes a computer vision-based system to address the pressing need for comprehensive home accident detection and response for the elderly. The system aims to fill the existing research gaps by expanding accident detection coverage and improving accident response procedures. The system receives the input real-time videos from a 2D RGB camera captured in indoor scenes, detects single-person accidents, stores the accident recordings, and sends alert messages. In the proposed framework, pose estimation is applied to extract the appearance feature, namely the elderly's skeleton key points from the input videos using MediaPipe Holistic. This feature is then input to the long short-term memory (LSTM) deep learning model to categorize the elderly's activities into falls, choking, and Activities of Daily Living (ADL) which indicate the absence of accidents. The framework is evaluated on UR Fall Detection (URFD) dataset and a self-generated dataset. The result shows that the framework achieved an accuracy of 87% in fall detection, 77% in choking detection, and 83% in ADL detection.